Blog
Field reports
Field notes from the middle of the agentic-coding shift: how I actually work with agents, what breaks, and what I change because of it. No expert claims here. I'd rather be wrong in public than fluent in a workflow that's already obsolete.
Transmission log11 entries
- Don't look at unreviewed codePrompts became loops, and the loop's most valuable step is the one most people skip: a review by a model from a different family, wired into the prompt itself, so findings get addressed before a human ever opens the diff.
- I'm not the expert, and that's the pointWhy I write about agentic coding without claiming to have it figured out: what I'm bad at (slowing down, evaluating, finishing), why I'd rather be wrong in public than fluent in a workflow that's already obsolete, and what you should expect from anything here.
- Orchestrate, don't operateThe most capable model is the most expensive one to let do grunt work. Put it in the orchestrator seat (planning, steering, judging subagent results) and hand the searching, reading, and typing to cheaper models.
- Turn the reasoning up before you turn the model downA smarter model needs less reasoning to reach the same answer, so it often costs less than a weaker model cranked to maximum effort. Adjust the effort dial first; change model tiers only when the dial runs out.
- Steal the idea, not the configEvery shared CLAUDE.md, skill, and prompt is someone else's answer to someone else's problem. Copying it wholesale imports their constraints and taxes every session. Start vanilla, and only add what a real, felt problem justifies.
- If your codebase is good for new-joiners, it's also good for agentsA fresh agent session is structurally a new joiner: empty context, no tribal knowledge, dependent on what's discoverable. The difference is that the new joiner pays the onboarding tax once, and your agent pays it several times a day.
- Maintainable for whom?Our whole quality vocabulary (maintainable, readable, clean) silently assumes a human will hand-edit the code later. If agents do the maintaining, the definition doesn't die, it inverts in surprising places.
- The adoption experts who won't adoptDevelopers spent a decade complaining that users won't adopt new software. Now agents arrive, and the same people go through denial, cherry-picked failure stories, and workflow-clinging themselves. The blocker isn't capability, it's identity.
- We need to get over GitCommits, branches, stashing, PRs: Git's ceremony is a coordination protocol for scarce human attention. Agents made work cheap and parallel, and suddenly the whole ritual feels like bureaucracy. Git the database survives; the ceremony has to go.
- Be aware of the context windowWhy Claude Code's auto-compaction quietly ruins long sessions, what context rot does to your results as the window fills up, and why the fix is awareness, not micro-managing every token.
- How to CLAUDE.mdA recent study measured what CLAUDE.md and AGENTS.md files actually do: developer-written ones help a little, LLM-generated ones hurt results and raise cost. Keep yours tight.